Molecular docking is now widely used for ligand discovery. The technique makes many approximations and though it has had noteworthy successes, it is neither fully reliable nor do we understand why it fails when it does. Here we develop well-controlled experimental systems to investigate specific problems in docking. These results inform development of new docking algorithms. The specific aims are: 1. Well-controlled model experimental systems to test docking algorithms. Extending work in the last period, we explore five new cavity sites to isolate specific terms in docking and molecular recognition. These cavities are entirely buried from solvent, dominated by single interactions, and bind multiple small ligands. We investigate the balance between ionic and polar interaction energies, desolvation, the role of bulk solvent, and the differences between recognition by aliphatic and aromatic groups. Predicted ligands are tested experimentally for binding and their geometries are determined by crystallography. These simple model systems only have impact if what we learn from them can be extended to biologically relevant sites. Also, some docking problems only emerge in such complicated sites. To investigate docking false negatives, we compare prospective docking and HTS campaigns targets. To investigate the role of library bias, we compare docking screens of fragment and drug-like libraries against two targets, -lactamase and 2-adrenergic receptor. 2. New docking algorithms. We focus on methods that can be directly tested in the experimental systems. We investigate: i. ligand and receptor conformational energy strain in docking, which are thought to be critical but have been poorly treated. Because the cavities are small, these terms are much more tractable than they are in larger sites for larger ligands. The cavities are also well suited to evaluating ii. ligand and receptor desolvation in docking, which may be specifically tested in the polar and ionic perturbations in the model cavity sites. We also use iii. thermodynamic integration to calculate ligand-cavity affinities, leveraging these studies to isolate key missing terms in docking. Finally, we return to iv. formal calculations of bias in screening libraries, and its role in docking.